GLANCE: continuous global-local exchange with consensus fusion for robust nodule segmentation

Accurate segmentation and detection of pulmonary nodules from computed tomography (CT) scans are critical for early lung cancer diagnosis but are hindered by the high diversity of nodule characteristics and the limitations of existing deep learning models. Conventional convolutional neural networks struggle with long-range context, while Transformers can neglect fine local details. We present GLANCE (Continuous Global-Local Exchange with Consensus Fusion), a novel dual-stream architecture designed to overcome these limitations. GLANCE features two parallel, co-evolving branches: a global context transformer to model long-range dependencies and a multi-receptive grouped atrous mixer to capture fine-grained local details. The core innovation is the cross-scale consensus fusion mechanism, which continuously integrates these complementary feature streams at every hierarchical scale, preventing representational clashes and promoting synergistic learning. A dual-head pyramid refinement decoder leverages these fused features to perform simultaneous nodule segmentation and center heatmap detection. Validated on four public benchmarks (LIDC-IDRI, LNDb, LUNA16, and Tianchi), GLANCE establishes a new state-of-the-art in both segmentation and detection. An extensive ablation study confirms that each architectural component, particularly the continuous fusion strategy, is critical to its superior performance.